x.comthreadGen AI Spotlight (@GenAISpotlight) on X
A tweet summarizing Anthropic's research finding that Z.ai's open-weight GLM-5.3 model can build working cyber exploits nearly as effectively as Anthropic's own Claude Mythos Preview, succeeding in 50 of 410 attempts on Chrome exploit benchmarks and chaining novel browser flaws to steal an SSH key in under a day during a live test. Anthropic warns that because GLM-5.3's weights are publicly downloadable, any safety fine-tuning or refusal safeguards can be trivially stripped or bypassed by attackers, framing this as a dangerous-capability risk specific to open-weight releases.
TL;DR Anthropic's capability eval finds Z.ai's open-weight GLM-5.3 nearly matches Claude at autonomously building working cyber exploits, and its downloadable weights make safeguards trivially bypassable.
Also seen from another angle🛡️ Security →
Technologies named here
Related in ai-safety
alignment.openai.comwriteup